ComfyUI Node
WanAnimatePlus ModelLoader
A ComfyUI node in WanAnimatePlus with 14 inputs and 1 output.
WanAnimatePlus ModelLoader
- compile_args
- block_swap_args
- lora
- vram_management_args
- extra_model
- fantasytalking_model
- multitalk_model
- fantasyportrait_model
- model
◄model▾►
◄base_precisionbf16►
◄quantizationdisabled►
◄load_deviceoffload_device►
◄attention_modesdpa►
◄rms_norm_functiondefault►
CategoryWanAnimatePlus
Inputs (14)
| Name | Type | Default | Description |
|---|---|---|---|
| model | COMBO | These models are loaded from the 'ComfyUI/models/diffusion_models' -folder | |
| base_precision | COMBO | bf16 | 4 options: fp32, bf16, fp16, fp16_fast |
| quantization | COMBO | disabled | Optional quantization method, 'disabled' acts as autoselect based by weights. Scaled modes only work with matching weights, _fast modes (fp8 matmul) require CUDA compute capability >= 8.9 (NVIDIA 4000 series and up), e4m3fn generally can not be torch.compiled on compute capability < 8.9 (3000 series and under) |
| load_device | COMBO | offload_device | Initial device to load the model to, NOT recommended with the larger models unless you have 48GB+ VRAM |
| attention_modeopt | COMBO | sdpa | 9 options: sdpa, flash_attn_2, flash_attn_3, sageattn, sageattn_3, radial_sage_attention, +3 |
| compile_argsopt | WANCOMPILEARGS | — | |
| block_swap_argsopt | BLOCKSWAPARGS | — | |
| loraopt | WANVIDLORA | — | |
| vram_management_argsopt | VRAM_MANAGEMENTARGS | Alternative offloading method from DiffSynth-Studio, more aggressive in reducing memory use than block swapping, but can be slower | |
| extra_modelopt | VACEPATH | Extra model to add to the main model, ie. VACE or MTV Crafter | |
| fantasytalking_modelopt | FANTASYTALKINGMODEL | FantasyTalking model https://github.com/Fantasy-AMAP | |
| multitalk_modelopt | MULTITALKMODEL | Multitalk model | |
| fantasyportrait_modelopt | FANTASYPORTRAITMODEL | FantasyPortrait model | |
| rms_norm_functionopt | COMBO | default | RMSNorm function to use, 'pytorch' is the new native torch RMSNorm, which is faster (when not using torch.compile mostly) but changes results slightly. 'default' is the original WanRMSNorm |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| model | WANVIDEOMODEL | — |